Upload CHANGELOG.md with huggingface_hub
Browse files- CHANGELOG.md +82 -0
CHANGELOG.md
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Changelog
|
| 2 |
+
|
| 3 |
+
All notable changes to the model artifacts in this repository.
|
| 4 |
+
|
| 5 |
+
Versions are tags. Deployments should pin a **Hugging Face commit SHA**, not a
|
| 6 |
+
tag and never `main` — see [README](README.md#usage).
|
| 7 |
+
|
| 8 |
+
## detector-2026.10.1
|
| 9 |
+
|
| 10 |
+
First release of the ONNX plate detection and text recognition artifacts.
|
| 11 |
+
|
| 12 |
+
### Detector — trained in this project
|
| 13 |
+
|
| 14 |
+
YOLOX-tiny, 640×640, single class, exported from `best_ckpt.pth` (epoch 93) with
|
| 15 |
+
torch 2.5.1+rocm6.2 at opset 11.
|
| 16 |
+
|
| 17 |
+
Trained on the Simanno annotated plate corpus (frame-grouped split with a
|
| 18 |
+
deliberate gap) plus 3,191 community-contributed images pseudo-labelled by an
|
| 19 |
+
8B Qwen3-VL teacher.
|
| 20 |
+
|
| 21 |
+
**Data corrections made before this release**, each because the label set was
|
| 22 |
+
measurably wrong rather than merely imperfect:
|
| 23 |
+
|
| 24 |
+
- A corrupt JPEG (`5e17d574-…`) that YOLOX's `cv2.imread` loader could not
|
| 25 |
+
decode. This aborted validation at epoch 10 in three separate training runs.
|
| 26 |
+
It had been previously reported as readable because PIL opens truncated JPEGs
|
| 27 |
+
without error; `cv2.imread` returns `None` for both missing and corrupt.
|
| 28 |
+
- 247 training images dropped because they contained a plate the teacher never
|
| 29 |
+
labelled. An unlabelled plate is false-negative supervision — the model is
|
| 30 |
+
penalised for detecting something that is genuinely there.
|
| 31 |
+
- 4 images of up to 108 megapixels excluded (decoding one costs ~324 MB RGB).
|
| 32 |
+
- 86 images excluded where EXIF orientation put the annotations in the wrong
|
| 33 |
+
coordinate frame for the training loader.
|
| 34 |
+
- Validation ground truth human-corrected: 43 plates the teacher omitted were
|
| 35 |
+
added and 12 misplaced boxes replaced, after review of all candidates.
|
| 36 |
+
- 1 leaked image removed from validation (the same photograph existed in both
|
| 37 |
+
the train and validation splits under different filenames).
|
| 38 |
+
|
| 39 |
+
### Measured performance
|
| 40 |
+
|
| 41 |
+
Split `v4` — 1,275 images, 1,143 plates, human-corrected ground truth,
|
| 42 |
+
detection IoU 0.3, confidence 0.3.
|
| 43 |
+
|
| 44 |
+
| | this release | previous model |
|
| 45 |
+
|---|---|---|
|
| 46 |
+
| recall | **0.926** | 0.906 |
|
| 47 |
+
| precision | **0.897** | 0.835 |
|
| 48 |
+
| false positives / image | **0.095** | 0.161 |
|
| 49 |
+
| recall, plates ≥100px | **0.958** | 0.899 |
|
| 50 |
+
| recall, plates <40px | 0.832 | 0.829 |
|
| 51 |
+
| latency mean / p95 | 20.8 / 38.6 ms | 21.7 / 39.6 ms |
|
| 52 |
+
|
| 53 |
+
Read against a measured seed-to-seed spread of **0.013 on recall, 0.013 on
|
| 54 |
+
precision, 0.021 on recall<40px**, the defensible claim is: **substantially
|
| 55 |
+
better precision and large-plate recall; recall overall and small-plate recall
|
| 56 |
+
unchanged.**
|
| 57 |
+
|
| 58 |
+
A previously reported 2-point *regression* on plates <40px was an artefact of
|
| 59 |
+
incomplete teacher labels and does not exist on corrected ground truth.
|
| 60 |
+
|
| 61 |
+
### Known limitations
|
| 62 |
+
|
| 63 |
+
- **Text accuracy is not measured.** The public corpus annotates plate boxes
|
| 64 |
+
and carries no transcriptions. Confidence tracks correctness monotonically in
|
| 65 |
+
a 20-plate pilot (CER 0.68 / 0.53 / 0.04 by confidence band), but 20
|
| 66 |
+
non-randomly-ordered plates are not a basis for setting a threshold, so
|
| 67 |
+
`PIPELINE_OCR_MIN_CONFIDENCE` remains 0.0.
|
| 68 |
+
- **Non-Latin recognition is unmeasured.** The recogniser runs with the
|
| 69 |
+
`alphanumeric` charset profile, which strips non-Latin characters by design.
|
| 70 |
+
- **The recogniser's upstream release is unrecorded.** See
|
| 71 |
+
[Provenance](README.md#provenance). Its dictionary is 18,383 characters, which
|
| 72 |
+
does not match PaddleOCR's stock `ppocr_keys_v1.txt` (6,622). Do not
|
| 73 |
+
regenerate it without first labelling non-Latin plates.
|
| 74 |
+
- Roughly 185 plates missed by **both** the teacher and the detector remain
|
| 75 |
+
absent from validation, so true recall is a range rather than a point.
|
| 76 |
+
|
| 77 |
+
### Rollback
|
| 78 |
+
|
| 79 |
+
`PIPELINE_BACKEND=llm` selects the external API path and is configuration-only —
|
| 80 |
+
no rebuild, no migration. The previous detector weights remain available on
|
| 81 |
+
request as `plate_yolox_tiny_640.corpus_v1.onnx`; they are not published here,
|
| 82 |
+
because this repository publishes exactly one supported detector version.
|